How to check an AI skill or MCP server is safe before you install it

A practical checklist for reviewing Claude skills, MCP servers and agent rules before you give them to your AI agent: what to read, red flags, and safer ways to try them.

Skills and MCP servers run with your agent's access. A skill can tell the agent to run commands; an MCP server is software running on your machine or a remote service. Most are fine — but it's worth five minutes of checking first.

1. Read the instructions yourself

Open the SKILL.md (every skill page here shows it in full). Look for:

Commands that download and run code (curl … | bash, npx from unknown packages).
Instructions to send files, keys or environment variables anywhere.
Text telling the agent to ignore your instructions or hide what it's doing.

If you can't tell what a command does, don't run it.

2. Check who made it

Prefer official sources (for example, the company whose product the MCP server connects to).
Look at the GitHub project: recent activity, open issues, number of contributors. Stars help but don't prove safety.

3. Give it only the access it needs

Use read-only API keys or tokens with the smallest scope.
For MCP servers, start with read-only tools and approve each tool call until you trust it.
Never paste production secrets into a skill file.

4. Try it somewhere safe first

Test new skills in a throwaway project or a container before using them on real work.

5. Keep it up to date — and remove what you don't use

Unused skills and servers are just extra risk.

How this directory helps

Entries shared by users are reviewed before they're published, and every entry links to its source so you can check it yourself. Browse curated skills or MCP servers, or learn what skills are.